[Dy2St][CINN] Use CINN as default backend (#71838)

SigureMo · Copilot · web-flow · commit f3ff2ffcd5d9 · 2025-04-10T10:54:32.000+08:00
---------

Co-authored-by: Copilot &lt;175728472+Copilot@users.noreply.github.com&gt;
diff --git a/python/paddle/jit/api.py b/python/paddle/jit/api.py
@@ -219,7 +219,7 @@ def to_static(
     function=None,
     input_spec=None,
     build_strategy=None,
-    backend=None,
+    backend="CINN",
     **kwargs,
 ):
     """
@@ -238,13 +238,18 @@ def to_static(
         build_strategy (BuildStrategy|None): This argument is used to compile the
             converted program with the specified options, such as operators' fusion
             in the computational graph and memory optimization during the execution
-            of the computational graph. For more information about build_strategy,
-            please refer to :code:`paddle.static.BuildStrategy`. The default is None.
-        backend(str, Optional): Specifies compilation backend, which can be `CINN` or
-            None. When backend is `CINN`, CINN compiler will be used to speed up
-            training and inference.
-        kwargs: Support keys including `property`, set `property` to True if the function
-            is python property.
+            of the computational graph. For more information about :attr:`build_strategy`,
+            please refer to :ref:`paddle.static.BuildStrategy <cn_api_paddle_static_BuildStrategy>`.
+            The default is ``None``.
+        backend(str, Optional): Specifies compilation backend, which can be ``"CINN"`` or
+            ``None``. When backend is ``"CINN"``, CINN compiler will be used to speed up
+            training and inference. default value is ``"CINN"``.
+        kwargs: Support keys including :attr:`property` and :attr:`full_graph`.
+
+          - property (bool): If True, the function will be treated as a property
+            function. The default is False.
+          - full_graph (bool): If True, the function will be converted into a
+            full static graph. The default is False.
 
     Returns:
         Tensor(s): containing the numerical result.
diff --git a/test/amp/test_amp_api.py b/test/amp/test_amp_api.py
@@ -456,7 +456,7 @@ def test_op_called_as_expected(self):
 
         func = SimpleModelIncludeSetValue()
         func = paddle.amp.decorate(func, level='O2')
-        func = paddle.jit.to_static(func, full_graph=True)
+        func = paddle.jit.to_static(func, full_graph=True, backend=None)
         input = paddle.randn((2, 3))
 
         with paddle.amp.auto_cast(level='O2', use_promote=False):
@@ -481,7 +481,7 @@ def test_pir_op_called_as_expected(self):
         with pir_dygraph_guard():
             func = SimpleModelIncludeSetValue()
             func = paddle.amp.decorate(func, level='O2')
-            func = paddle.jit.to_static(func, full_graph=True)
+            func = paddle.jit.to_static(func, full_graph=True, backend=None)
             input = paddle.randn((2, 3))
 
             paddle.amp.debugging.enable_operator_stats_collection()
diff --git a/test/dygraph_to_static/CMakeLists.txt b/test/dygraph_to_static/CMakeLists.txt
@@ -4,7 +4,7 @@ file(
   "test_*.py")
 string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
 set(SOT_ENVS SOT_LOG_LEVEL=0 MIN_GRAPH_SIZE=0 STRICT_MODE=False
-             SOT_ENABLE_STRICT_GUARD_CHECK=True)
+             SOT_ENABLE_STRICT_GUARD_CHECK=True ENABLE_CINN_IN_DY2ST=False)
 
 if(WIN32 AND NOT WITH_GPU)
   # disable on Windows CPU CI for timeout
diff --git a/test/legacy_test/test_jit_layer.py b/test/legacy_test/test_jit_layer.py
@@ -33,31 +33,34 @@
 paddle.seed(1)
 
 
-class Net(paddle.nn.Layer):
-    def __init__(self):
-        super().__init__()
-        self.fc1 = paddle.nn.Linear(4, 4)
-        self.fc2 = paddle.nn.Linear(4, 4)
-        self._bias = 0.4
-
-    @paddle.jit.to_static(
-        input_spec=[InputSpec([None, 4], dtype='float32')], full_graph=True
-    )
-    def forward(self, x):
-        out = self.fc1(x)
-        out = self.fc2(out)
-        out = paddle.nn.functional.relu(out)
-        out = paddle.mean(out)
-        return out
-
-    @paddle.jit.to_static(
-        input_spec=[InputSpec([None, 4], dtype='float32')], full_graph=True
-    )
-    def infer(self, input):
-        out = self.fc2(input)
-        out = out + self._bias
-        out = paddle.mean(out)
-        return out
+def create_net():
+    class Net(paddle.nn.Layer):
+        def __init__(self):
+            super().__init__()
+            self.fc1 = paddle.nn.Linear(4, 4)
+            self.fc2 = paddle.nn.Linear(4, 4)
+            self._bias = 0.4
+
+        @paddle.jit.to_static(
+            input_spec=[InputSpec([None, 4], dtype='float32')], full_graph=True
+        )
+        def forward(self, x):
+            out = self.fc1(x)
+            out = self.fc2(out)
+            out = paddle.nn.functional.relu(out)
+            out = paddle.mean(out)
+            return out
+
+        @paddle.jit.to_static(
+            input_spec=[InputSpec([None, 4], dtype='float32')], full_graph=True
+        )
+        def infer(self, input):
+            out = self.fc2(input)
+            out = out + self._bias
+            out = paddle.mean(out)
+            return out
+
+    return Net()
 
 
 class TestMultiLoad(unittest.TestCase):
@@ -70,7 +73,7 @@ def tearDown(self):
     def test_multi_load(self):
         paddle.disable_static()
         x = paddle.full([2, 4], 2)
-        model = Net()
+        model = create_net()
         with enable_to_static_guard(False):
             forward_out1 = model.forward(x)
             infer_out1 = model.infer(x)
@@ -88,7 +91,7 @@ def test_multi_load(self):
 
     def test_multi_jit_load(self):
         x = paddle.full([2, 4], 2)
-        model = Net()
+        model = create_net()
         with enable_to_static_guard(False):
             forward_out1 = model.forward(x)
             infer_out1 = model.infer(x)
@@ -102,18 +105,21 @@ def test_multi_jit_load(self):
         np.testing.assert_allclose(infer_out1, infer_out2, rtol=1e-05)
 
 
-class SaveLinear(paddle.nn.Layer):
-    def __init__(self):
-        super().__init__()
-        self.linear = paddle.nn.Linear(80, 80)
+def create_save_linear():
+    class SaveLinear(paddle.nn.Layer):
+        def __init__(self):
+            super().__init__()
+            self.linear = paddle.nn.Linear(80, 80)
 
-    @paddle.jit.to_static(
-        input_spec=[InputSpec(shape=[None, 80], dtype='float32')],
-        full_graph=True,
-    )
-    def forward(self, x):
-        out = self.linear(x)
-        return out
+        @paddle.jit.to_static(
+            input_spec=[InputSpec(shape=[None, 80], dtype='float32')],
+            full_graph=True,
+        )
+        def forward(self, x):
+            out = self.linear(x)
+            return out
+
+    return SaveLinear()
 
 
 class TestMKLOutput(unittest.TestCase):
@@ -126,7 +132,7 @@ def tearDown(self):
     def test_mkl_output(self):
         paddle.disable_static()
         with _dygraph_place_guard(place=paddle.CPUPlace()):
-            net = SaveLinear()
+            net = create_save_linear()
             model_path = os.path.join(self.temp_dir.name, 'save_linear')
             paddle.jit.save(net, model_path, combine_params=True)
 
@@ -139,7 +145,7 @@ def test_mkl_output(self):
 
     def test_mkl_jit_output(self):
         with _dygraph_place_guard(place=paddle.CPUPlace()):
-            net = SaveLinear()
+            net = create_save_linear()
             x = paddle.ones([498, 80])
             orig_out = net.forward(x)
             model_path = os.path.join(self.temp_dir.name, 'save_linear')
diff --git a/test/legacy_test/test_zero_size.py b/test/legacy_test/test_zero_size.py
@@ -177,7 +177,7 @@ def forward_func(x):
         for place in self.places:
             paddle.device.set_device(place)
             static_forward_func = paddle.jit.to_static(
-                forward_func, full_graph=True
+                forward_func, full_graph=True, backend=None
             )
 
             for dtype in self.dtypes:
@@ -308,7 +308,7 @@ def gradient_func(x):
                     self.assertEqual(x.data_ptr(), 0)
 
                     static_gradient_func = paddle.jit.to_static(
-                        gradient_func, full_graph=True
+                        gradient_func, full_graph=True, backend=None
                     )
                     (x_grad,) = static_gradient_func(x)
 
@@ -440,7 +440,7 @@ def gradient_func(x):
                     self.assertEqual(x.data_ptr(), 0)
 
                     static_gradient_func = paddle.jit.to_static(
-                        gradient_func, full_graph=True
+                        gradient_func, full_graph=True, backend=None
                     )
                     (x_grad,) = static_gradient_func(x)
 
diff --git a/test/prim/pir_prim/test_auto_recompute_dy2static.py b/test/prim/pir_prim/test_auto_recompute_dy2static.py
@@ -94,7 +94,7 @@ def product_rms_norm_inputs(self, place):
     def cal_rms_norm_res(self, place):
         weight, hidden = self.product_rms_norm_inputs(place)
         net = PrimNet()
-        net = paddle.jit.to_static(net, full_graph=True)
+        net = paddle.jit.to_static(net, full_graph=True, backend=None)
         program = net.forward.get_concrete_program(weight, hidden)[
             -1
         ].program.program
diff --git a/test/prim/pir_prim/test_decomp_fallback.py b/test/prim/pir_prim/test_decomp_fallback.py
@@ -38,7 +38,9 @@ def setUp(self):
         self.p = paddle.to_tensor(0.0, dtype=self.dtype)
 
     def test_fallback(self):
-        static_func = paddle.jit.to_static(self.func_api, full_graph=True)
+        static_func = paddle.jit.to_static(
+            self.func_api, full_graph=True, backend=None
+        )
         dynamic_func = self.func_api
 
         out = static_func(self.x, self.p)
@@ -62,7 +64,9 @@ def setUp(self):
         self.value = paddle.to_tensor(2, dtype=self.dtype)
 
     def test_fallback(self):
-        static_func = paddle.jit.to_static(self.func_api, full_graph=True)
+        static_func = paddle.jit.to_static(
+            self.func_api, full_graph=True, backend=None
+        )
         dynamic_func = self.func_api
 
         out = static_func(self.x, self.value)
@@ -86,7 +90,9 @@ def setUp(self):
         self.axis = paddle.to_tensor(1, dtype=paddle.int64)
 
     def test_fallback(self):
-        static_func = paddle.jit.to_static(self.func_api, full_graph=True)
+        static_func = paddle.jit.to_static(
+            self.func_api, full_graph=True, backend=None
+        )
         dynamic_func = self.func_api
 
         out = static_func(self.x, self.axis)
@@ -110,7 +116,9 @@ def setUp(self):
         self.axis = paddle.to_tensor([0, 2], dtype=paddle.int64)
 
     def test_fallback(self):
-        static_func = paddle.jit.to_static(self.func_api, full_graph=True)
+        static_func = paddle.jit.to_static(
+            self.func_api, full_graph=True, backend=None
+        )
         dynamic_func = self.func_api
 
         out = static_func(self.x, self.axis)
@@ -136,7 +144,9 @@ def setUp(self):
         self.axis = [0]
 
     def test_fallback(self):
-        static_func = paddle.jit.to_static(self.func_api, full_graph=True)
+        static_func = paddle.jit.to_static(
+            self.func_api, full_graph=True, backend=None
+        )
         dynamic_func = self.func_api
 
         out = static_func(self.x, self.axis)
